ci: add vLLM GPU smoke workflow with dynamic EC2 runner lifecycle (#541)

[skamenan7]

Summary

Adds a manual GPU smoke workflow for #541. It runs the existing stateless Responses API test through Praxis against vLLM on an ephemeral AWS GPU runner.

Flow

1. Start runner: GitHub Actions uses OIDC to assume the Praxis AWS role and starts an ephemeral g4dn.xlarge runner.
2. Run smoke test: The runner verifies the GPU, starts the pinned vLLM image with Qwen/Qwen3-0.6B, starts OGX, builds Praxis, waits for both services, and runs test_stateless_request.
3. Clean up: A final job always terminates the EC2 runner, including when the test fails.

Safety and validation

• Manual dispatch only.
• AWS access uses GitHub OIDC and the repository-scoped role from aaet-infra MR !33.
• The vLLM GPU workflow uses an image pinned by digest.
• Failure logs are captured and the GPU runner is cleaned up.
• Action lint, shellcheck, Rust build, and test collection pass locally.

The first live GPU run is still pending access to the configured repository variables and runner-registration secret.

[praxis-bot-app]

Unsigned commits: b57a227. Please sign your commits.